Privacy for complex data

Learned decisions need to be based on accurate information. Decision support systems and data-driven models can help decision makers and planners to make these decisions.

Nevertheless, data-driven models need to use data of good quality for their training and these data are not always available or easy to access. Data is sensitive and models can contain traces of these data.

Membership inference attacks are examples of attacks to ML models to prove that some particular records have been used when training the model.

Attribute inference attacks and model inversion attacks are other examples of attacks that can reveal sensitive information about individuals from the models. A large number of data protection mechanisms have been developed for standard databases publishing. That is, when data is of the form of an SQL database with records described in terms of variables or attributes.

Nevertheless, data privacy mechanisms have limitations when there are temporal aspects in data, and when there are dependencies between objects. It is difficult to provide multiple data releases from a time-varying data set. Similarly, it is difficult to protect objects that share information or correlated features.

In this project we will focus on these two problems: temporal dependencies and interrelated components. We are interested in data publishing for complex data. In particular, we will provide solutions for data publishing for dynamic graphs and metering data.